sample_embedding_folder2/0694032.txt

Ticket Name: TDA2: TIDL Eltwise Layer output not matching with caffe output

Query Text:
Part Number: TDA2 Hi, I have Convolutional layers, InnerProduct and Eltwise layers in my model. The caffe ouput and tidl outputs of the convolutional layer and innerproduct layers are matching. But in the elwise layer, some of the elements of the tidl matrix vary too much with the elements of caffe output. For example, let's say Eltwise layer adds two blobs of size 1x5. Here is how the caffe and tidl output looks like: Caffe output: [22.24 -2.33 -15.85 -14.89 -11.05] TIDL output: [21.29 59.05 -15.00 -14.03 -10.16] As you could see the [1] element vary too much. To convert the data to floating point, I am dividing the output data by the OutQ in the logs as mentioned in this thread(e2e.ti.com/.../2370081 Regards, Navinprashath.R.R

Responses:
Please find the prototxt and caffe model file. The tidl and caffe outputs of the second Eltwise layer are not matching. Also the logs during the import and the input files(lenet_48_48.y) are also attached. PS: The weights here not used in the actual model. But the similar error can be reproduced with this network. Regards, Navinprashath.R.R Not able to upload file because of size attaching drive link. https://drive.google.com/file/d/19Zt7PlV5-yIpD-inALszIwhuxHDUwrPz/view?usp=sharing

Hi, The problem occurs when the input to the Eltwise layer is of opposite signs. Regards, Navinprashath.R.R

Hi Naveen, We don't support this combination. Thanks, Praveen

Hi Praveen, Thanks Praveen. Could we expect this combination support in future release? It is for the same network I am not able to increase the numParamBits to 12. Is it related to this issue? Regards, Navinprashath.R.R

There is no plan to add support for this combination now.